Battery swapping stations (BSSs) have rapidly developed over the past decade due to their ability to alleviate the issues of long charging times and limited travel distances per charge for electric scooters. However, BSSs face the challenge of high charging cost, which hinders their further development. To identify the root causes, in this paper, we analyze a battery-swapping dataset consisting of 41 BSSs and more than 7600 batteries. We find that under Time-of-Use (TOU) pricing, the charging strategy of BSSs and user behavior lead to a high share (e.g., 51.2%) of electricity usage during peak price periods, thereby increasing the charging cost. We then provide evidence for potential opportunities to reduce the charging cost. Inspired by these findings, we design a BSS charging scheduling system called CharTD to reduce BSS charging cost. The system aims to optimize the charging strategy under the TOU pricing mechanism while considering real-time user demand. Specifically, we formulate the charging scheduling problem as a nonconvex-constrained nonlinear optimization problem. We then design a deep reinforcement learning-based approach to solve the charging scheduling problem. To verify the effectiveness of the system, we implement and evaluate CharTD with the above battery-swapping operation dataset. The experimental results show that CharTD effectively reduces the charging cost of BSSs by more than 13.3% compared with three baseline methods.